OpenAI has introduced GPT-5.6 Sol, Terra, and Luna with a rollout that highlights both their capabilities and access limitations. The new family of models arrives with improvements in software engineering, computing use, professional work, scientific research, and cybersecurity, but during the preview, it will only be available to a select group of trusted organizations via API and Codex. ChatGPT is not included in this initial phase.
The notable aspect isn’t just OpenAI’s staged deployment, which is common for frontier models. What’s significant is that the company itself acknowledges sharing its plans and capabilities with the U.S. government before launch, and that, in coordination with the administration, they are beginning with a limited group of partners whose participation has been communicated to the government. OpenAI also states that it does not believe this kind of government-access process should become standard in the long term.
For a tech outlet, the core issue isn’t merely whether GPT-5.6 Sol surpasses competitors in capability. Underlying that is the realization that access to the most advanced AI is increasingly resembling operating strategic infrastructure subject to national security, compliance, and political approval rather than simply requesting an API.
Sol, Terra, and Luna: three models, a narrow access gateway
OpenAI categorizes GPT-5.6 into three tiers. Sol is the flagship model, Terra aims for a balance between capacity and cost, and Luna is presented as the fastest and most economical option. The company also introduces a new reasoning mode max for Sol and an ultra mode that employs coordinated sub-agents for complex tasks.
Pricing-wise, Sol costs $5 per million input tokens and $30 per million output tokens. Terra’s prices drop to $2.50 and $15, respectively, while Luna is priced at $1 and $6. OpenAI also enhances prompt caching with explicit cutoff points and a minimum cache life of 30 minutes, with cached reads maintaining a 90% discount.
| Model | Approach | Input | Output |
|---|---|---|---|
| GPT-5.6 Sol | Maximum capacity | $5/M tokens | $30/M tokens |
| GPT-5.6 Terra | Balance | $2.50/M tokens | $15/M tokens |
| GPT-5.6 Luna | Speed and cost | $1/M tokens | $6/M tokens |
The issue is that these prices only matter if you can actually use the model. During the preview, there’s no public request system, no waitlist, and OpenAI Support cannot manually add an organization. Access depends on selected organizations already working with an OpenAI account representative.
This limitation changes the competitive landscape. A company might have the budget, technical team, use case, and a real need for frontier AI, but not necessarily access. In traditional cloud setups, the barrier was paying for capacity and adhering to terms of service. In this AI phase, the barrier may be simply being on a list.
Cybersecurity, biology, and the argument for containment
OpenAI justifies phased deployment through the combination of increased capability and the need for safeguards. Sol is described as its most powerful model to date, with enhancements in agentic capabilities, programming, biology, and cybersecurity. Regarding security, the company asserts that the model improves functions like code review, vulnerability research, patch development, debugging, education, and defensive testing, while attempting to hinder prohibited offensive uses.
The company states that GPT-5.6 Sol does not cross the “Cyber Critical” threshold of its Preparedness Framework. According to OpenAI, assessments with Chromium and Firefox identified bugs and exploitation primitives, but the model did not autonomously generate a fully functional chain under the tested conditions. This nuance is important: it doesn’t eliminate risk but illustrates how OpenAI aims to position the model within a controlled deployment framework.
There are also new real-time checks for dual-use areas, especially biology and cybersecurity. OpenAI explains that some requests may be blocked or delayed while additional controls are executed, and that these mechanisms could occasionally affect legitimate requests.
The technical reasoning is sound. A more capable model can significantly assist defenders, researchers, and security teams, but it can also lower barriers for malicious actors. When the response to this dilemma is a closed, coordinated list with a government, the conversation shifts from product security alone to broader strategic considerations.
Open source is no longer just a lab rarity
The other half of the debate is outside OpenAI. As frontier proprietary models grow more powerful and controlled, open alternatives have closed the gap. DeepSeek V3.2, for example, is licensed under MIT for assets distributed in open repositories, including weights and code. GLM-5.2 from Z.ai is also available under MIT, with no regional restrictions. The Qwen family maintains open licenses like Apache 2.0 in official repositories.
It’s important not to oversimplify. “Open model” doesn’t always mean complete transparency. Often, weights and code are published, but not all training data or the entire process. Not all open models match the top proprietary ones across all tasks. However, for a tech company, value isn’t just about benchmarks; it’s about being able to deploy, audit, adapt, and maintain alternatives without relying on external approval.
This is where practical differences emerge. A closed model may deliver better performance, support, integration, and managed security. An open model offers operational control, enabling deployment on in-house infrastructure, policy adjustments, working without sending data externally, and reducing dependence on unilateral access, pricing, or jurisdiction changes.
For many organizations, the smart strategy isn’t choosing sides but designing an architecture that combines proprietary and open models, measures performance per task, and avoids reliance on a single provider as a political point of failure.
AI as a geopolitical layer of software
The launch of GPT-5.6 Sol arrives at a point where AI is no longer just a productivity tool. It begins to influence cybersecurity, scientific research, software automation, industrial design, defense, biotech, and critical business operations. The closer it gets to these domains, the more it resembles a regulated technology.
This shift will have consequences. U.S. providers will face pressure from their government. Chinese providers will respond to their own political and industrial frameworks. Europe must decide whether to settle for consuming third-party models or develop enough capacity to avoid being caught between blocs.
For technical teams, the conclusion is less ideological and more architectural: it’s no longer enough to ask “which model responds best.” Instead, questions about access control, deployment location, policy changes, migration options, data exfiltration, and local maintenance become central.
GPT-5.6 Sol may be a remarkable technical advance, but its limited preview reveals a reality many preferred to ignore: the frontier of AI won’t be a neutral API for everyone equally. It will be a layer of infrastructure governed by rules, interests, jurisdictions, and restrictions.
This is why open source is gaining significance in the debate. Not necessarily because it’s always superior, but because it offers an option that becomes strategic when access to powerful models depends on external decisions: the ability to operate without permission.
Frequently Asked Questions
What is GPT-5.6 Sol?
It’s the flagship model of OpenAI’s new GPT-5.6 family, alongside Terra and Luna. OpenAI presents it as their most capable model yet.
Is it available on ChatGPT?
Not during the initial preview. OpenAI states that GPT-5.6 is only accessible via API and Codex for a limited group of partners and organizations.
Can a company request public access?
No. During this phase, there is no public form, waitlist, or manual support registration. OpenAI contacts selected organizations directly.
Why is the U.S. government’s role important?
Because OpenAI acknowledges coordinating this preview with the U.S. government and sharing initial group participation with the administration.
Does this make open source more relevant?
Yes, though it doesn’t always replace proprietary models. Open models offer control, self-deployment, and continuity when organizations prefer not to fully depend on closed APIs.
